Re: [RFC] iovec in ->aio_read/->aio_write

[Shailabh Nagar <nagar@watson.ibm.com>]

Janet Morgan wrote:

> Here's a patch for aio readv/writev support.  Basically it adds:
> 
> - two new opcodes (IOCB_CMD_PREADV and IOCB_CMD_PWRITEV)
> - a field to the iocb for the user vector
> - aio_readv/writev methods to the file_operations structure

I presume f_op->aio_readv could point to __generic_file_aio_read for most
filesystems.

Would f_op->aio_writev need a new wrapper function for 2.5.42 ?
f_op->aio_write eventually calls generic_file_write which uses a different inode
from generic_file_writev. So f_op->aio_writev might need to point to a function
like generic_file_writev but using the same inode as generic_file_write.


> - routine aio.c/io_readv_writev, which borrows heavily from do_readv_writev. 
> 
> I tested this using the aio dio patch that Badari submitted a while back. 
> I compared:
>                 readv/writev io_submit for a vector of N iovecs 
>                 vs read/write io_submit for N iocbs.
> 
> My performance data is only preliminary at this point, but aio readv/writev 
> appears to outperform aio read/write -- twice as fast in some cases.  The 
> results generally make sense to me:  while there is only one io_submit in both 
> cases, aio readv/writev shortens codepath (one instead of N calls to the 
> underlying filesystem routine) and should normally result in fewer 

Twice as fast looks good !

> bios/callbacks (at least for direct-io).  As importantly, aio readv/writev 
> in my testing also reduces the number of (system) calls to io_getevents.

It would be interesting to see the performance boost when <iov length> events
are retrieved at once, using the min_nr parameter of io_getevents.


--Shailabh